AI beats goalkeepers at predicting which way penalty taker will shoot

Recent research has shown that deep learning models, trained on a dataset of over 1000 penalty kicks from football matches, outperform real-life goalkeepers in predicting the direction of the ball. This breakthrough could have significant implications for the sport.

David Freire-Obregón, a researcher at the University of Las Palmas de Gran Canaria, Spain, led the study. He explains, “Penalty kicks are crucial moments in soccer, often determining the outcome of important games. Despite this, goalkeepers mainly rely on intuition during these high-pressure situations. We wanted to investigate whether machine learning could provide valuable insights into shot direction based on a kicker’s body motion.”

The study involved analyzing 1010 penalty kicks from televised matches in Spain. Of these, 640 kicks were suitable for AI analysis, while the rest were excluded due to factors like blurriness or obstruction. Each kick was inputted into 22 deep learning models, which had to predict whether the ball would go left, right, or down the middle based on video footage and the kicker’s dominant foot.

The most successful model achieved a 52% accuracy rate in determining the ball’s direction, surpassing the 46% accuracy rate of real goalkeepers in the matches. When the middle option was removed, the model’s accuracy rose to 64%, almost 10 percentage points higher than human goalkeepers with the same information.

Freire-Obregón notes the significance of subtle motion cues in revealing a player’s intent before kicking the ball. While this information could benefit goalkeepers in training sessions, implementing AI predictions in real-time match scenarios presents challenges.

Looking ahead, the researchers plan to explore whether these predictive probabilities can be identified in advance during penalty kicks, solely based on the kicker’s pre-shot motion. The goal is to determine how early such predictions can be made while maintaining a high level of accuracy.